The problem with the added KernelState (which shouldn't be wrapped in a Ref) is that GPUCompiler passes it as a value and not a reference, see https://github.com/JuliaGPU/GPUCompiler.jl/blob/32b4fc87eeece6302dd47cf20e255ee510acfc4a/src/irgen.jl#L498-L504, to work around performance regressions introduced by passing it as a byval pointer, see JuliaGPU/CUDA.jl#1167. So if we want to extend KernelState support to back-ends that expect a pointer, we'll have to revert JuliaGPU/GPUCompiler.jl@5eb6098, make sure it doesn't re-introduce said performance regressions (presumably by not relying on NVPTX' byval handling but by manually converting it to values like we did before). This is a little tricky, because the expansion of byval needs optimization to clean-up, so KernelState argument insertion will need to be moved earlier in the pipeline. I think this should take a day or so, and I should have some time to look at this somewhere in the coming weeks.

JuliaGPU/GPUCompiler.jl#715 is a possible fix for this.

Replaced by #657